Computer Vision-Based Wildfire Smoke Detection Using UAVs
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Hindawi Mathematical Problems in Engineering Volume 2021, Article ID 9977939, 9 pages https://doi.org/10.1155/2021/9977939 Research Article Computer Vision-Based Wildfire Smoke Detection Using UAVs Ehab Ur Rahman ,1 Muhammad Asghar Khan ,2 Fahad Algarni ,3 Yihong Zhang,1 M. Irfan Uddin ,4 Insaf Ullah,2 and Hafiz Ishfaq Ahmad5 1 College of Information Science & Technology, Donghua University, Shanghai, China 2 Hamdard Institute of Engineering & Technology, Islamabad 44000, Pakistan 3 College of Computing and Information Technology, University of Bisha, Bisha, Saudi Arabia 4 Institute of Computing, Kohat University of Science and Technology, Kohat, Pakistan 5 Faculty of Engineering, UniversitiTeknologi Malaysia, Johor, Malaysia Correspondence should be addressed to M. Irfan Uddin; irfanuddin@kust.edu.pk Received 21 March 2021; Revised 14 April 2021; Accepted 20 April 2021; Published 29 April 2021 Academic Editor: Dr. Dilbag Singh Copyright © 2021 Ehab Ur Rahman et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. This paper presents a new methodology based on texture and color for the detection and monitoring of different sources of forest fire smoke using unmanned aerial vehicles (UAVs). A novel dataset has been gathered comprised of thin smoke and dense smoke generated from the dry leaves on the floor of the forest, which is a source of igniting forest fires. A classification task has been done by training a feature extractor to check the feasibility of the proposed dataset. A meta-architecture is trained above the feature extractor to check the dataset viability for smoke detection and tracking. Results have been obtained by implementing the proposed methodology on forest fire smoke images, smoke videos taken on a stand by the camera, and real-time UAV footages. A microaverage F1-score of 0.865 has been achieved with different test videos. An F1-score of 0.870 has been achieved on real UAV footage of wildfire smoke. The structural similarity index has been used to show some of the difficulties encountered in smoke detection, along with examples. 1. Introduction monitoring tools for the last couple of years [4–8]. A high- definition and lightweight cameras can generate an aerial Wildfire is a colossal threat to damaging the human and photograph with specific location information when con- wildlife ecosystem. Statistics show that wildfires in Northern nected to UAVs along with global positioning systems California in the United States caused more than 40 deaths (GPSs) [9]. Besides, cost-effectively, a well-organized swarm and about 50 missing individuals in 2015 [1–3]. There were of UAVs can easily accomplish a complex task. some major wildfire outbreaks in several countries around Image recognition attained a state-of-the-art perfor- the world in the year 2019. It was seen to be the most mance using deep convolutional neural networks (DCNNs); unfortunate year for such incidents. Moreover, 3500 square its architecture and learning scheme leads to an effective miles of the Amazon rainforest have been burnt down by extractor of sophisticated, high-level features that are highly wildfires. A forest fire recently caused 89 fatalities in Aus- robust to input transformations [10]. However, imple- tralia and burned 3500 homes. It became of such incidents of menting deep learning and computer vision techniques in great importance to detect wildfires accurately in advance the application for wildfire smoke detection is scarce. Be- when it turns into chaos. Traditional methods of wildfire sides, the limitations and difficulties in such kinds of detection, which are mainly based on human observation techniques are not widely discussed. Object detectors mainly from watchtowers, are inefficient. The inefficiency is pri- based on video fire detection methods can be categorized marily due to the spatiotemporal connection. Unmanned into two classes, i.e., flame detection and smoke detection. aerial vehicles (UAVs) have been extensively used as Since the smoke generated by forest fires is observable before
2 Mathematical Problems in Engineering the flames, video smoke detection acquires more attention smoke videos taken by a UAV and smoke images with for early fire alarm in forest fire protection engineering. The different kinds of backgrounds and lighting conditions. Also, traditional video smoke detection methods mainly em- some of the limitations and difficulties are discussed found phasize the combination of static and dynamic features for during the study of this research. smoke detection. The typical features of smoke contain color, texture, motion orientation, and so on [11]. These 1.2. Organization of the Paper. The other sections of the different characteristics can get better performances in paper are presented as follows. Section 2 discusses the specific images dataset [12]. However, due to the poor ro- different datasets followed by the training of the objector bustness of algorithms, the performances incline to be detector in Section 3. Section 4 presents details on the results unfavorable in different images dataset, and those ap- and concludes the work, and Section 5 concludes the entirety proaches can barely remove sophisticated interference in of the paper. real engineering applications. Currently, object detection achieved a lot of progress due to the use of GA, PSO, ANN, and DCNNs [13–18]. Modern 2. Material and Methods object detectors founded on these networks—such as Faster 2.1. Real Smoke Training Images. The dataset used to train R-CNN [19], SSD [20], and YOLOv3 [21]—are now robust the model consists of 14096 images of real smoke, both enough to be deployed in customer products (e.g., Google comprised of thin and dense smoke. The smoke is typically Photos and Pinterest Visual Search), and some are fast generated in different scenarios. One set of smoke is gen- enough to be run on mobile devices such as MobileNet. Most erated from burning dry leaves and small bushes, which is of these object detectors are deployed in different applica- one of the fuel causes of igniting a forest fire and smoke. The tions [22]. However, it can be challenging for practitioners to images are shown in Figures 1(a) and 1(b) as dense and thin select what architecture is more appropriate to their ap- smoke, respectively. Another set of images consists of smoke plication. Standard accuracy metrics do not clarify the entire images taken in different light conditions such as yellow and options, such as mean average precision (mAP); for practical white light as it affects smoke color and texture, as shown in deployment of computer vision systems, running time and Figures 1(c) and 1(d). These images have been taken from memory usage are also important. For example, mobile [11], which is also comprised of some other images having devices in many cases need a small memory footprint, and added smoke to the forest background. Examples of these self-driving cars need real-time executions. SSD achieves a images are presented in Figures 1(e) and 1(f ). A third set good trade-off between speed and precision. SSD runs a consists of smoke images taken from various open sources convolutional network on input image a single time and from the Internet that present real-time emergencies such as calculates a feature map following a small 3 × 3 sized con- an apartment on fire or a vehicle. Such images are shown in volutional kernel on this feature map to predict the Figures 1(g) and 1(h). bounding boxes and categorization probability. SSD also uses anchor boxes on various aspect ratios similar to Faster- RCNN and learns the counterbalance to a definite extent 2.2. Test Images. The proposed methodology and object than learning the box. To handle the scale, SSD predicts detector are evaluated based on test smoke images to check bounding boxes after multiple convolutional layers. Since the generalization ability of the trained object detection every convolutional layer function at a diverse scale, it can model. These images are a collection of smoke images taken detect objects of varying scales [23]. from the UAV with a camera of 12 MP [11], and also some of There are several ways of comparing images as if they are them are taken from phone cameras specifically using an identical or near-identical such as structure similarity index iPhone 6 s camera with a 12 MP rear camera, while the rest measure (SSIM), mean square error (MSE), normalized are selected from the open-source datasets available on the color histogram, and local sensitivity hashing. These Internet. Figure 2 illustrates such kinds of images. methods have various benefits over one another. SSIM tries to model the modification of the image’s structural details. 2.3. Test Videos. With static camera videos as well as forest SSIM is more robust capable of disclosing changes in the smoke videos in real-time taken by UAV, the performance of image structure rather than just the perceived change [24]. various object detection models is also tested. The focus of this work is on choosing an object detector that for real-time object detection that has a better trade-off between speed 1.1. Contributions. In this work, we present a dataset, and precision. Table 1 displays the technical specifications of grouping several images from different sources such as thin, the UAV, along with the UAV image (see Figure 3) [25], dense with different color, and texture smoke images, taken which can be used to capture the smoke images. from different scenarios such as wildfire and other emer- gency conditions such as building fires and fires from an 3. Training and Detection explosion. The SSD Inception-V2 state-of-the-art models are trained, and their different parameters such as dropout, DCNNs have presently dominated computer vision tasks in batch normalization, and learning rate are tuned to choose which region-based object detection methods are state-of- the best model for real-time fire detection in videos. the-art. These methods have different advantages, such as Comparisons of the results are obtained on several wildfire removing the gruesome work of manual feature extraction.
Mathematical Problems in Engineering 3 (a) (b) (c) (d) (e) (f) (g) (h) Figure 1: Training images taken from different sources. Figure 2: Test images from different sources. The network learns patterns from the images without provides a good precision in detection objects of different needing any preprocessing. Currently, several different ar- sizes as compared with Faster-RCNN architecture. chitectures of feature extractors are available. Selecting one In this work, a pretrained feature extractor has been for a specific application is a trivial subject. According to trained by transfer learning with a custom classifier with two [26], different speed versus accuracy comparison has been fully connected layers and a final log softmax classifier to presented. Figure 4 illustrates such speed versus accuracy classify the proposed dataset into two classes with one class trade-offs between the current state-of-the-art object de- of smoke including both dense and thin smoke images and tection models. Speed and accuracy are both of keen im- another class having fire images. The classification task was portance for real-time smoke detection. From Figure 4, it is aimed to check the feasibility of the proposed dataset. The clear that single shot detectors (SSDs) achieve a better trade- promising results have been presented in the results section off in the aspect of swiftness and accuracy. Also, SSD of this paper, along with some examples. SSD as a meta-
4 Mathematical Problems in Engineering Table 1: Technical parameters of the UAV. First, the feature extractor Inception-V2 pretrained on Technical parameters Value the COCO dataset is trained by transfer learning using the proposed dataset with the object detector SSD for making Camera lens FOV 81.9° 25 mm predictions of smoke. The trained model is then tested with Take-off weight 300 g Video resolution FHD: 1920 × 1080 30 p different test videos and images to check the feasibility of the Max hovering time 15 minutes model. Max flight speed 31 mph Classification results of the wildfire smoke classifier Endurance 16 minutes model trained by using the dataset comprising of both Positioning system GPS/GLONASS smoke and fire images are presented in Table 2. Metrics used Obstacle sensing range 11–16 ft(0.2–5 m) for evaluation are presented in the study by Bashir and Operating temperature Porikli [27] and Forman and Scholz [28]. The metrics are 320 to 1040 F range True positive (TP), False Positive (FP), True Negative (TN), Transmitter power(EIRP) 2.4 GHz False Negative (FN), and False alarm rate (FAR) along with Max transmission distance 100 m (distance), 50 m (height) detection rate recall precision and F-score. Such kind of Dimensions 143 × 143 × 55 mm 2.412–2.462 GHz; metrics are currently used in the computer vision com- Operating frequency munity for evaluation object detection models. Another 5.745–5.825 GHz Camera sensor ½.3” CMOS; effective pixels: 12MP metric that is used for comparing the similarity in the ISO range Video: 100–3200 photo: 100–1600 structure formation of two images is SSIM [24]. Formulas of Electronic shutter speed 2–1/8000 s the metrics are presented as follows: Max vidoe bitrate 24 mbps TP Photo format JPEG recall(Re) � , (1) Video format MP4 (TP + FN) Controllable range Pitch: –850 to 00 Stabilization 2-axis mechanical (pitch, roll) TP precision (Pr) � , (2) Velocity range 36 kph at 2 m above ground (TP + FP) Altitude range 0–8 m Operating range 0–30 m 2 · (Pr · Re) F − score � , (3) (Pr + Re) architecture and Inception-V2 as a feature extractor have been chosen to be more suitable for real-time smoke de- FP false alarm rate(FAR) � , (4) tection as they offer better speed versus accuracy trade-off, as (TP + FP) shown in Figure 4 taken from the study by Huang [26]. In the proposed work, different images taken from dif- (TP + TN) accuracy � , (5) ferent sources are collected together to increase the richness of FN the training data primarily focusing on images having thin smoke as it serves as an alarm before a fire starts, and there is TP tracker detection rate (TDR) � , an immense need to detect the smoke at this starting stage to TG prevent the ignition and spreading of the wildfire. Before (6) testing the proposed model, we trained the model on more than 14000 training image samples comprised of both dense 2µx µy + c1 2σ xy + c2 smoke and thin smoke and smoke in a different light, color, SSIM(x, y) � 2 2 2 2 . (7) µx + µy + c1 σ x + σ y + c2 and texture as well as different backgrounds. For validation of the model, more than 3100 images were used. This approach Equation (7) presents a comparison of two windows, i.e., was aimed to improve the generalization ability of the smoke small subsamples despite the whole image, leading to a better detection model. We then tested the model using wildfire approach that can sense for changes in the structure of the smoke images taken from a drone, mobile phone camera, and image. The parameters of equation (7) confine the (x, y) lo- some from Internet open sources which present real-time cation of the N × N window in each image, the mean of the scenarios of both forest fire and in day-to-day life emergencies pixel intensities in the x and y direction, and also the variance of so that the proposed approach can be used in any kind of intensities in the x and y direction, along with the covariance. situation for future projects. Test videos taken from standby cameras of both thin and 4. Research Findings and Discussion dense smoke with different backgrounds, lighting condi- tions, and from different distances are used for inferencing. As introduced in Section 3, we trained a classifier on the Also, real footage of wildfire and smoke taken by a drone is smoke image dataset along with another set of fire images. tested. Results have been presented in Section 4 for analysis, Some of the classification results are shown in Figure 4 along along with some discussions on the limitations and diffi- with top k class probabilities. Figure 6(a) presents a test culties found in the research of this study of wildfire smoke image taken from a thin smoke dataset, Figure 6(b) presents detection. The overall workflow of the experiments is pre- a test image taken from dense smoke images, and Figure 6(c) sented in Figure 5. presents a test image from a fire image.
Mathematical Problems in Engineering 5 Battery compartment Nut Nose direction Aircraft nose mark Propeller Direction LED red Motor Direction LED red Direction LED green Direction LED green LED indicator Adhesive tape Receiver antenna Landing gear Camera mounting frame Compass Figure 3: Main on-board components of the UAV (quadcopter) [25]. 40 35 30 Overall mAP 25 20 15 10 0 200 400 600 800 1000 GPU time Feature extractor Inception resnet V2 MobileNet Inception V2 Resnet 101 Inception V3 VGG Figure 4: Speed vs. accuracy trade-off. Single-shot Wildfire Input images detector smoke CNN Inception-V2 Testing UAV-camera Training Validation Transfer learning MS-COCO Images Video dataset frames Results Alarming Human system observation Figure 5: Diagram of the proposed methodology for smoke detection.
6 Mathematical Problems in Engineering Table 2: Two class classification results of the classifier. Test set Samples TP FP TN FN FAR Recall Precision F-score Smoke 20 18 2 0 0 0.1 1.00 0.90 0.947 Fire 20 18 0 2 0 0 1.00 1.00 1.00 Smoke Smoke Fire 0 0 0 100 100 100 200 200 200 0 100 200 0 100 200 0 100 200 Smoke Smoke Fire Fire Fire Smoke 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 (a) (b) (c) Figure 6: Classification result of feature extractor with high-class probabilities. Table 2 presents obtained results of the classification of feasibility of the dataset on thin smoke also. Another good the dataset into two classes, i.e., fire and smoke. The results result is from the drone footage sample which is the main from Table 2 show that the feature extractor is generalizing focus of this study to detect smoke in such kind of scenario well and is learning different smoke patterns, i.e., thin, dense, achieved 83.2% accuracy and an F-score of 0.870 which white, and so on. This small experiment was aimed to prove proves the suitability of this study. The results from Table 3 the viability of using the dataset in the training of the object show that this approach is feasible to implement in real-time detector for real-time smoke detection. applications. The trained wildfire smoke detector model, i.e., SSD Figure 8 presents the mean average recall (mAR) and Inception-V2 has been tested with images taken from a UAV mean average precision (mAP) of each test dataset that has shown in Figure 7(a) along with the detection score and been evaluated with the wildfire smoke detection model. bounding boxes. Figure 7(b) presents the frames captured These metrics are famous for evaluating the overall gener- from video of thin smoke and dense smoke generated by alization ability of object detection models. The respective burning dry leaves and shrubs. These videos are recorded on figure proves that our smoke detection model efficiently a mobile phone camera. Frames captured of different in- generalized to different datasets which comprise distinctive stances from real-time UAV footage of wildfire and smoke image sets with different properties such as thin smoke and are presented in Figure 7(c) with respective bounding boxes. dense smoke, and an image of smoke taken from different Figure 7(e) presents a frame comprising of both dense and angles and approaches. thin smoke along with detections. Some of the limitations have been observed while The test results are presented in Table 3. It is observed acquiring the results. Figure 9(a) shows smoke due to that Video 2, i.e., a dense smoke video has the highest wildfire, and the frame is captured from real footage. The F-score among all that is because of the rich texture, shape, texture and color of this image are nearly similar to the and color. The lowest F-score is observed from thin video texture and color of the cloud shown in Figure 9(b). Such a samples and that is because of light color and features kind of coincidence makes it difficult for the object de- captured by the camera, but still, the model achieves an tector to differentiate between them. The structural accuracy of 64% and an F-score of 0.747, proving the similarity between the two images has been calculated
Mathematical Problems in Engineering 7 (a) (b) (c) (d) (e) Figure 7: Sample’s result taken from different test videos along with detection boxes. Table 3: Results of real smoke video samples. Smoke Nonsmoke Test set TP FP TN FN FAR Accuracy Detection rate/recall Precision F1-score sample samples Images 74 30 50 21 32 1 0.296 0.766 0.980 0.704 0.818 Video 1 (dense smoke near) 816 0 599 216 0 1 0.265 0.734 0.998 0.735 0.847 Video 2 (dense smoke far) 307 0 301 6 0 0 0.019 0.980 1.000 0.980 0.989 Video 3 (thin + dense 313 0 245 68 0 0 0.217 0.782 1.000 0.783 0.878 smoke) Video 4 (drone footage) 3983 1479 3099 884 1447 32 0.221 0.832 0.989 0.778 0.870 Video 5 (thin smoke) 308 0 184 124 0 0 0.402 0.600 1.000 0.597 0.747 Video 6 (thin smoke) 289 0 185 104 0 0 0.359 0.640 1.000 0.640 0.780 mAR mAP 1.2 1.2 Mean average precision Mean average recall 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 Images Dense smoke ... Dense smoke ... Thin + dense ... Drone ... Thin smoke 1 Thin smoke 2 Images Dense smoke ... Dense smoke ... Thin + dense ... Drone ... Thin smoke 1 Thin smoke 2 Figure 8: Mean average recall (mAR) and mean average precision (mAP) for different test datasets. using the structure similarity index measure (SSIM). The night in the dark is also tested with the trained object value of SSIM calculated for these images shows that they detector. Though the smoke detector does detect some of have a 63% structure similarity. Both of them have been the frames, the overall accuracy is unsatisfactory. This was detected as smoke by the trained object detector. The meant to give intuition to the readers for such kinds of confidence score and bounding box are shown in difficulties, which may be addressed in future research by Figure 9(c). Also, in Figure 9(d), a video of the fire taken at designing new approaches.
8 Mathematical Problems in Engineering (a) (b) (c) (d) Figure 9: Smoke vs. cloud similarity comparison with SSIM value. 5. Conclusion References In this paper, SSD Inception-V2 was chosen to be a viable [1] A. D. Syphard, V. C. Radeloff, N. S. Keuler et al., “Predicting detector of wildfire smoke in videos taken by UAVs both spatial patterns of fire on a southern California landscape,” in terms of accuracy and speed. Different smoke image International Journal of Wildland Fire, vol. 17, no. 5, datasets such as one generated by using a synthetic process pp. 602–613, 2008. [2] N. Nauslar, J. Abatzoglou, and P. Marsh, “The 2017 north Bay and another from real smoke images are used to train the and southern California fires: a case study,” Fire, vol. 1, no. 1, model. One of the significant solutions is presented to p. 18, 2018. detect thin and dense smoke in videos taken by UAVs as [3] T. John and A. Abatzoglou, “Impact of anthropogenic climate previous methods comprise images or static camera change on wildfire across western US forests,” Park Williams videos. The test results promise of extending the solution Proceedings of the National Academy of Sciences, vol. 113, to real-time drone surveillance. An F1-score of 0.784 and no. 42, pp. 11770–11775, 2016. 0.747 has been achieved on test videos of thin smoke [4] I. Colomina and P. Molina, “Unmanned aerial systems for surpassing the previous literature. Limitations and diffi- photogrammetry and remote sensing: a review,” ISPRS culties found in the study are discussed along with an Journal of Photogrammetry and Remote Sensing, vol. 92, example using structural similarity index as a quantifying pp. 79–97, 2014. parameter. In the future, the proposed solution can be [5] S. S. V. Vijayakumar, C. S. Kumar, V. Priya, L. Ravi, and V. Subramaniyaswamy, “Unmanned aerial vehicle (UAV) extended to detect smoke in real-time UAV footage in based forest fire detection and monitoring for reducing false different light and weather conditions along with de- alarms in forest-fires,” Computer Communications, vol. 149, signing a fire alarm. The performance of the model on thin pp. 1–16, 2020. smoke can be further improved by enriching the thin [6] F. Noor, M. A. Khan, A. Al-Zahrani, I. Ullah, and K. A. Al- smoke image dataset mainly taken by UAVs in different Dhlan, “A review on communications perspective of flying weather and light conditions. ad-hoc networks: key enabling wireless technologies, appli- cations, challenges and open research topics,” Drones, vol. 4, Data Availability no. 65, pp. 1–14, 2020. [7] B. Alzahrani, O. S. Oubbati, A. Bernawi, A. Atiquzzaman, and The data used to support the findings of this study are D. Alghazzawi, “UAV assistance paradigm: state-of-the-art in available from the corresponding author upon request. applications and challenges,” Journal of Network and Com- puter Applications, vol. 166, no. 102706, pp. 1–44, 2020. [8] M. A. Khan, I. M. Qureshi, and F. A. Khanzada, “Hybrid Conflicts of Interest communication scheme for efficient and low-cost deployment of future flying ad-hoc network (FANET),” Drones, vol. 3, The authors declare that they have no conflicts of interest. no. 16, pp. 1–20, 2019.
Mathematical Problems in Engineering 9 [9] X. Li, Y. Zhao, J. Zhang, and Y. Dong, “A hybrid PSO al- [25] “General thoughts on putting together your own quadcopter”, gorithm based flight path optimization for multiple agricul- Available online: https://unleashthebot.com/best-drone-kits/, tural UAVs,” in Proceedings of 2016 IEEE 28th International 2021. Conference on Tools with Artificial Intelligence (ICTAI), San [26] J. Huang, “Speed/accuracy trade-offs for modern convolu- Jose, CA, USA, November 2016. tional object detectors,” in Proceedings of the IEEE Conference [10] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature on Computer Vision and Pattern Recognition, Honolulu, HI, hierarchies for accurate object detection and semantic seg- USA, June 2017. mentation,” in Proceedings of the IEEE Conference on Com- [27] F. Bashir and F. Porikli, “Performance evaluation of object puter Vision and Pattern Recognition, Colombus, OH, USA, detection and tracking systems,” in Proceedings 9th IEEE June 2014. International Workshop on PETS, New York, NY, USA, June [11] Q.-x. Zhang, G.-h. Lin, Y.-m. Zhang, G. Xu, and J.-j. Wang, 2006. “Wildland forest fire smoke detection based on faster R-CNN [28] G. Forman and M. Scholz, “Apples-to-apples in cross-vali- using synthetic smoke images,” Procedia Engineering, vol. 211, dation studies,” Acm Sigkdd Explorations Newsletter, vol. 12, pp. 441–446, 2018. no. 1, pp. 49–57, 2010. [12] E. U. Rahman, Y. Zhang, S. Ahmad, H. I. Ahmad, and S. Jobaer, “Autonomous vision-based primary distribution systems porcelain insulators inspection using UAVs,” Sensors, vol. 21, no. 3, p. 974, 2021. [13] Z. Ali and T. Mahmood, “Complex neutrosophic generalised dice similarity measures and their application to decision making,” CAAI Transactions on Intelligence Technology, vol. 5, no. 2, pp. 78–87, 2020. [14] T. Sangeetha and G. M. Amalanathan, “Outlier detection in neutrosophic sets by using rough entropy based weighted density method,” CAAI Transactions on Intelligence Tech- nology, vol. 5, no. 2, 2020. [15] C. Zhu, W. Yan, X. Cai, S. Liu, T. H. Li, and G. Li, “Neural saliency algorithm guide bi-directional visual perception style transfer,” CAAI Transactions on Intelligence Technology, vol. 5, no. 1, pp. 1–8, 2020. [16] C.-F. J. Kuo, J.-M. Liu, M. L. Umar, W.-L. Lan, C.-Y. Huang, and S.-S. Syu, “The photovoltaic-thermal system parameter optimization design and practical verification,” Energy Con- version and Management, vol. 180, pp. 358–371, 2019. [17] M. Safa, M. Ahmadi, J. Mehrmashadi et al., “Selection of the most influential parameters on vectorial crystal growth of highly oriented vertically aligned carbon nanotubes by adaptive neuro-fuzzy technique,” International Journal of Hydromechatronics, vol. 3, no. 3, p. 238, 2020. [18] B. R. Murlidhar, R. K. Sinha, E. T. Mohamad, R. Sonkar, and M. Khorami, “The effects of particle swarm optimisation and genetic algorithm on ANN results in predicting pile bearing capacity,” International Journal of Hydromechatronics, vol. 3, no. 1, p. 69, 2020. [19] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: towards real-time object detection with region proposal networks,” in Proceedings of Advances in Neural Information Processing Systems, Montreal, Canada, December 2015. [20] W. Liu, D. Anguelov, D. Erhan et al., “Ssd: single shot multibox detector,” in Proceedings of European Conference on Computer Vision, Amsterdam, Netherlands, October 2016. [21] H. Shah, YOLO Vs. SSD: Choice of a Precise Object Detection Method, https://technostacks.com/blog/yolo-vs-ssd, 2020. [22] E. U. Rahman, Y. Zhang, S. Ahmad, H. I. Ahmad, and S. Jobaer, “Autonomous vision-based primary distribution systems porcelain insulators inspection using UAVs,” Engi- neering, 2020, preprint. [23] H. Shah, What is the Main Difference Between Yolo and Ssd?, Technostacks Infotech Pvt. Ltd., Ahmedabad, India, 2018, https://technostacks.com/blog/yolo-vs-ssd/. [24] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004.
You can also read
